Coverage overview
Every surface AgentKeeper can collect from, what each one sees, and whether it can block before an action runs.
A surface is one source of evidence. Each installs independently and appears in Settings → Connectors, where an admin can turn its ingestion on or off.
What each surface covers
| Surface | Sees | Can block | Installs on |
|---|---|---|---|
| Runtime hooks | Tool calls, shell commands, file access, prompts | Yes, before execution | Workstation, inside the agent |
| Agent Runtime SDK | Events from agents you build (TypeScript, Python) | Yes, in your own code | Your application |
| MCP Gateway | Every MCP tool call and its arguments | Yes, at the gateway | Workstation, as a local proxy |
| Browser extension | Prompts in web AI surfaces | Yes, before submit | Chrome, Edge, Firefox |
| Claude Enterprise | Governed Claude prompts across claude.ai, Cowork, Claude Code | Yes, org-wide checkpoint | Nothing — Anthropic sends to you |
| Claude OTLP | Post-execution telemetry, token cost, detector findings | No | Managed settings or env vars |
| eBPF runtime sensor | AI processes, network connections, container and Kubernetes context | No | Linux hosts |
| Endpoint discovery (EDR) | Which machines have AI agents installed | No | Nothing — read-only API sync |
| Anthropic Compliance API | Anthropic Enterprise activity | No | Nothing — API sync |
| OpenAI Compliance API | ChatGPT Enterprise audit logs | No | Nothing — API sync |
| Microsoft Agent 365 | Enterprise agent registry, ownership, risk | No | Nothing — API sync |
Blocking requires the surface to be asked before the action runs. Everything else is after-the-fact evidence.
Which ones you need
Runtime hooks alone answer "what did the agent do on this machine". Add the MCP Gateway when agents reach external tools, and the browser extension when people paste code into web AI.
| If you want to | Install |
|---|---|
| Stop dangerous commands in the IDE | Runtime hooks |
| Control which MCP servers agents can reach | MCP Gateway |
| Stop source code leaving through a chat window | Browser extension |
| Find shadow AI on machines you do not manage | Endpoint discovery (EDR) |
| Audit AI use in SaaS, without touching endpoints | Compliance API connectors |
| See AI workloads in containers and CI | eBPF runtime sensor |
Plan gating
| Surface | Minimum plan |
|---|---|
| Runtime hooks, MCP Gateway, browser extension | Free |
| Endpoint discovery (EDR) | Team |
| Claude Enterprise, Compliance APIs, OTLP forwarding | Enterprise |
| Single sign-on, directory sync | Enterprise |
Turning a surface off
Disabling a connector stops ingestion — it does not uninstall anything. Installed hooks keep running and their events are dropped. Evidence already stored stays visible. Uninstall from the workstation with Uninstall or the fleet rollback procedure.